NBER WORKING PAPER SERIES

LANGUAGE ENVIRONMENT AND MATERNAL EXPECTATIONS:

AN EVALUATION OF THE LENA START PROGRAM

Flavio Cunha
Marsha Gerdes
Qinyou Hu
Snejana Nihtianova

Working Paper 30837

http://www.nber.org/papers/w30837

NATIONAL BUREAU OF ECONOMIC RESEARCH

1050 Massachusetts Avenue
Cambridge, MA 02138
January 2023

We thank Delitza Hernandez, Michelle Gifford, Shanae Smith, and Mary Triguba for their
excellent research assistance. We are thankful to all of the LENA Foundation team. Jill Gilkerson
and Steve Hannon for all of the support and encouragement relating to the research aspects of this
project. Traci Martin and Jess Simmons supported our implementation of the LENA Start
Program at the Children’s Hospital of Philadelphia. Their dedication to this program, which we
have heard from several other national sites, energized our team. We are grateful to Jeffrey
Richards, who generously shared crucial knowledge on operationalizing the LENA system data
and provided valuable comments about this research. We thank participants at the following
conferences: Sociedade Brasileira de Econometria in Natal, Brazil; CONBATRI in Juiz de Fora,
Brazil; the Austin-Bergen Conference at UT Austin; and the briq Belief Conference at the
University of Bonn. Finally, we thank seminar participants at the Federal Reserve Bank in New
York City, EESP at FGV in Sao Paulo, Boston University, the University of Minnesota, the
CEPA Seminar at Stanford University, Vancouver School of Economics, and Rochester
University. This research was supported by grant 1R01HD073221-01A1 by the National Institute
of Health. All remaining errors are ours. The views expressed herein are those of the authors and
do not necessarily reflect the views of the National Bureau of Economic Research.

NBER working papers are circulated for discussion and comment purposes. They have not been
peer-reviewed or been subject to the review by the NBER Board of Directors that accompanies
official NBER publications.

© 2023 by Flavio Cunha, Marsha Gerdes, Qinyou Hu, and Snejana Nihtianova. All rights
reserved. Short sections of text, not to exceed two paragraphs, may be quoted without explicit
permission provided that full credit, including © notice, is given to the source.
Language Environment and Maternal Expectations: An Evaluation of the LENA Start Program
Flavio Cunha, Marsha Gerdes, Qinyou Hu, and Snejana Nihtianova
NBER Working Paper No. 30837
January 2023
JEL No. I21,I24

ABSTRACT

Research documents that parental beliefs influence early investments in children, which, in turn,
determine early human capital and, eventually, other skills children acquire in later stages of the
lifecycle, such as literacy. Our paper reports the results of an experimental evaluation of the
LENA Start Program, a group- and center-based parenting program that teaches the science of
early language development, models verbal interaction behaviors with children, and provides
objective feedback to improve the early language environment. The intervention changes parental
beliefs and impacts the quantity and quality of parental linguistic input.

Flavio Cunha Qinyou Hu

Department of Economics Department of Economics
Rice University Rice University
P.O. Box 1892 P.O. Box 1892
Houston, TX 77251-1892 Houston, TX 77251-1892
and NBER qinyou.hu@rice.edu
flaviocunha@rice.edu
Snejana Nihtianova
Marsha Gerdes Rice University
Children's Hospital of Philadelphia P.O. Box 1892
3615 Civic Center Blvd Houston, Texa 77251-1892
Philadelphia, PA 19104 snejana.nihtianova@rice.edu
gerdes@chop.edu

A online appendix is available at http://www.nber.org/data-appendix/w30837

I. Introduction

The 2017 National Assessment of Educational Progress reports a sizeable

socioeconomic gap in reading skills. In the United States, 46% of low-income1 children enrolled
in fourth grade have reading skills below the basic level. In contrast, the same rate for children
raised in middle- and high-income households is about 18%. Reading is not just critical for
school success. Research shows that the capacity to read proficiently confers positive economic
returns in the labor market (e.g., Hanushek et al., 2015; Murnane et al., 2000; Vignoles et al.,
2011). Beyond financial returns, research also shows that the ability to comprehend written text
is associated with better access to health services (e.g., Dexter et al., 1998; Sandiford et al.,
1995), improvements in self-esteem (Bown, 1990; Canieso-Doronila, 1996), higher levels of
empowerment (Burchfield et al., 2002; Kagitcibasi et al., 2005) and political participation (Egbo,
2000; Hannum & Buchmann, 2003; Ireland, 1994; Purcell-Gates & Waterman, 2000), promotion
of cultural change, and cultural diversity preservation (Carr-Hill et al., 2001; Chebanne & Nyati-
Ramahobo, 2003; Norwood, 2003; Robinson-Pant, 2000; Tarawa, 2003). Reading proficiency is
so fundamental in modern life that it is considered a basic human right (UNESCO, 1975).

Theoretically, reading proficiency requires two distinct skills: decoding and language
comprehension skills (Gough and Turner, 1986; Scarborough et al., 2009).2 Empirically, most
children who cannot read proficiently fail to do so because they lack appropriate language skills
(Foorman, Petscher, and Herrera, 2018, OECD, 2016). In addition, a large body of research has
documented steep socioeconomic (SES) gradients in language development observed before
the school years (Fernald et al., 2012; Hart & Risley, 1995; Hoff, 2006; Huttenlocher et al.,
2010).

Developing these critical language skills requires greater exposure to language so that
children can practice and hone their language processing skills and simultaneously increase
their vocabulary (e.g., Weisleder and Fernald, 2013; Pan et al., 2005; Gilkerson et al., 2018;
Romeo et al., 2018). This literature finds a sizeable socioeconomic gradient in the language
environment. The relation between family SES and the child’s early language skills is partly due
to the quantity and quality of parent speech directed towards the child during day-to-day
interactions (e.g., Hart and Risley, 1995; Hoff, 2003, 2006; Rowe and Goldin-Meadow, 2009;
Fernald et al., 2012).

1
Low-income children are children eligible for the National School Lunch Program.
2
We present a detailed literature review in Appendix A.

1
 Recent literature shows that low-income mothers underestimate the returns on early
investments (e.g., Cunha et al., 2013; Attanasio et al., 2019). In addition, heterogeneity in
parental beliefs predicts early parental investments (e.g., Attanasio et al., 2019; Cunha et al.,
2022; Biroli et al., 2022). Indeed, Bhalotra et al. (2020) show that once they control for
perceived returns and costs of early investments (i.e., breastfeeding), there is little role for
heterogeneity in preferences. In addition, this research also finds sizeable socioeconomic
gradients in parental beliefs (Boneva and Rauh, 2018; List et al., 2021).

Low expectations about the returns to investments might rationalize the existence of
parenting interventions in the developed (e.g., the Nurse-Family Partnership in the USA, see
Olds et al., 2004; Heckman et al., 2017) and developing worlds (e.g., the Jamaica Home Visiting
Program, Grantham-McGregor et al., 1991; and its replications in other countries such as the
ones described in Attanasio et al., 2014; Carneiro et al., 2019; and Heckman et al., 2020). In
parallel, developmental psychologists have created parent-directed interventions to improve
early language environments (Leech et al., 2018, McGillion et al., 2017a; Suskind et al., 2015;
Rowe and Leech, 2019; and Ramirez et al., 2020). These studies show that interventions that
combine education (i.e., teaching parents about the science of brain and language
development), coaching (i.e., modeling language interaction behaviors to parents), and
objective feedback (i.e., providing objective statistics about the quality of the language
environment) might produce impressive impacts on language outcomes. These interventions
might work by elevating parental expectations about the returns to early investments.

This paper reports the results of evaluating a scalable intervention designed to improve
disadvantaged children’s language environment, known as the LENA Start Program. The LENA
Start Program occurred within the Philadelphia Human Development longitudinal study
(henceforth, PHD Study). This study recruited 822 mothers during the second trimester of their
first pregnancy and elicited mothers’ subjective beliefs about the importance of parent-child
interaction for child development. When the children were nine to twelve months old, the study
visited the families in their homes and measured family investments. In addition, the study
assessed cognitive and language development when the children were two years old.

In the PHD Study data, the language environment’s heterogeneity – measured by

conversational turn counts – has a fragile association with family income. Our research
hypothesizes that heterogeneity in parental beliefs drives parental linguistic input gaps.
Therefore, in this paper, we investigate if it is possible to change parental linguistic input and, if
so, whether parental beliefs are one of the mechanisms of this change.

2
 The LENA Start Program is ideal for testing our hypothesis because of its curriculum. It
explains information about the importance of the language environment for language
development, offers a menu of natural and practical actions that parents can take to improve
their child’s language environment, and provides objective feedback about the quality of the
child’s language environment for the duration of the intervention. These activities occur weekly
for thirteen weeks. Unlike typical parenting programs, LENA Start is a low-cost group-centered
model that may help parents build social capital and a solid social support network.3,4

For the LENA Start Program evaluation, we identified a group of 289 low-income families
whose children’s ages were appropriate for the intervention. We randomly allocated half of the
families and children to the control group. We assigned the other half to receive an invitation to
participate in the LENA Start Program. We successfully located 94 (or 65%) of the 145 parents
in the control group and 95 (or 66%) of the 144 parents in the invitation group. Once we found
the parents, we tried to consent the parents to participate in the study activities. Nearly 75% (71
out of 94) of parents in the control group agreed to participate in the study. Approximately 68%
(65 out of 95) of the treatment-group parents decided to participate in the study. Therefore, we
finally enrolled 136 families in the LENA Start evaluation study. Due to the group format, our
intervention suffers from an issue of non-random attrition. We admit that non-random attrition
can compromise random assignment. However, Section II. A. and Appendix C show that our
sample is balanced conditional on the random assignment and that the random assignment
indicator strongly predicts the final attendance status. Appendix F shows that the treatment
modality does not drive our sample attrition. The results are robust to attrition correction using
inverse probability weighting Wooldridge (2010) or the Lee (2009) treatment effects bounds.

The evaluation study’s primary outcomes are objective measures of the language
environment, which we obtained with the LENA System (Gilkerson and Richards, 2008). This
system allows researchers to measure several dimensions of the language environment: adult-
child conversational turns, the number of adult words spoken around the child, and the amount
of exposure to TV or other electronics. To complement the LENA System, one of the
innovations of our measures is that we also use the LENA Advanced-Data Extractor (ADEX) to
break the audio file into segments that include conversations between the focus child and an
adult (male or female). The LENA System estimate does not distinguish between words directed

3
Zhang et al. (2015) and Gilkerson et al. (2018) are interventions that use linguistic feedback to parents.
4
As of Spring 2019, the LENA Start Program is being implemented by hospital systems, school districts,
public libraries, and NGOs that serve parents and their young children. For a list of current LENA Start
program sites, see https://www.lena.org/about/#where-is-lena.

3
to the child and words overheard by the child. However, with LENA ADEX, we can identify who
initiated the conversation (the child or a male or a female adult) and the number of
conversational turns per segment as categorized by the segment initiator (the child, a female
adult, or a male adult). In that way, we can investigate whether children or adults initiate more
conversations (audio segments).

We also collected survey measures at both baseline end endline on four potential
mechanisms that could explain any positive findings of the impact of the LENA Start Program on
parent-child linguistic interaction. Specifically, we measured (1) parental beliefs and (2) parental
knowledge about the importance of the language environment for a child’s language
development. We also measured (3) parental self-efficacy, defined as the parent’s belief in their
ability to perform the parenting role competently, and (4) their sense of social support
perception. We followed the instrument design of Cunha et al. (2013) to measure parental
beliefs. First, we presented mothers with two hypothetical scenarios of language environment
(“investments”) and asked them to predict the child’s future language development. Next, we
used the instrument Suskind et al. (2016) developed to measure parental knowledge. Then, we
followed Bandura (1977) to measure parental self-efficacy. Finally, we used four items to
measure their social support perception due to the group nature of this particular intervention.

We analyze the program’s Intent to Treat (ITT) and Local Average Treatment effects
(LATE). We find strong evidence that the LENA Start Program significantly impacts
conversational turns and adult words in conversations between the focus child and an adult.
Controlling for conversational turns at baseline, we find the LENA Start Program significantly
increases conversational turns between adults and the focus child by 31.4% of a standard
deviation. With the measures from LENA ADEX, we find that the differences in conversational
turns arise because the LENA Start Program children initiate a more significant number of
conversations – ITT estimates indicate that the treatment group children begin nearly 40% of a
standard deviation more conversational turns than their control counterparts. The Program also
significantly increases adult words spoken to the child. The significant increase is the product of
a higher number of conversational turns initiated by children and responded to by parents in
ways that do not stop the conversation but encourage the children to continue to talk and
interact. Finally, we find the Program does not impact exposure to TV or other electronics.

Exploring the mechanisms, we find that the change in parental beliefs about the
language environment’s importance for language development explains the program’s impacts.
Indeed, our mediation analysis shows that changing parental beliefs can explain 14 percent of

4
the total program impact. We do not find evidence supporting other mechanisms we test with
our data.

Our paper relates to the literature on parenting programs for language development in
developmental psychology. Leech et al. (2018), McGillion et al. (2017), Suskind et al. (2013),
Suskind et al. (2015), and Rowe and Leech (2019) have designed and evaluated parenting
education programs that target malleable aspects of the parental language input. These
interventions translate scientific information about how parental language input (a critical
component of a child’s early language environment) predicts early language development. Also,
these interventions employ parental coaching based on objective measures of the child’s
language environment. As discussed below, these ingredients are present in the LENA Start
Program. Our innovation is that we evaluate not a prototypical program but a program that
school districts, public libraries, and children’s hospitals have adopted nationally. In fact,
according to the LENA Foundation, nearly 2,000 families participate in the LENA Start Program
annually.

Our paper also contributes to the literature to the growing literature that investigates the
malleability of parental beliefs. For example, Attanasio et al. (2019) found that a home-visiting
program in Colombia did not shift parental beliefs, which is consistent with the result that the
same intervention did not produce permanent changes in family investments. In contrast,
Carneiro et al. (2019) evaluated a large-scale, low-cost parenting program and found that the
intervention raised beliefs and improved children’s early environments. Finally, List et al. (2021)
showed that an intervention combining education and feedback delivered through home visits
increases beliefs and parent-child interactions. Our work contributes to this literature by showing
that low-cost, center- and group-based interventions that combine education, coaching, and
objective feedback improve children’s early environments through the malleability of parental
beliefs.

We organize the rest of this paper in the following way. Section II describes our study
procedures and details of the LENA Start Program, including participants, recruitment,
intervention, and measurement. We also present the empirical methodology for evaluating the
program’s effects. Section III presents results, including the estimates of the impact of the LENA
Start Program on statistics of the language environment, and describes our findings of the
potential mechanisms of the Program. Section IV concludes.

5
II. Study Procedures

This section describes the study procedures approved by the University of Pennsylvania
and Children’s Hospital of Philadelphia Institutional Review Boards.

A. Participants

We used three criteria to determine LENA Start Program Evaluation Study eligibility.
First, we restricted eligibility to PHD Study mothers who resided in the inner-city and whose
children were at most 33 months by May/2017 because the LENA Start Program is designed for
children up to that age. 5 Additionally, we attempted to recruit mothers who authorized the
research team to be contacted for participation in future PHD Studies. We identified a group of
289 mothers who satisfied all of the three inclusion criteria.

After we identified the eligible families and before we recruited them, we grouped PHD
Study Participants according to the child’s date of birth. We created ten blocks with 26 to 28
mothers in each group. Next, we took a random draw from a uniform distribution for each
mother in each group. We created an “ordered invitation list” by ordering the mothers in
descending order of the draw from the uniform distribution within each block. Let ܼ௜ ൌ 1 if
mother ݅ is at the top half of the invitation list and ܼ௜ ൌ 0 if the mother is at the bottom half of the
list. In what follows, we refer to the mothers at the top half of the list (i.e., those with ܼ௜ ൌ 1) as
participants randomly assigned to the “treatment group” while mothers at the bottom half (those
with ܼ௜ ൌ 0) as study participants randomly assigned to the “control group.” Figure 1 describes
the results of the randomization procedure. We randomly assigned 145 of the 289 eligible
mothers to the control group and the remaining 144 to the group invited to participate in the
LENA Start Program.

The group format of the LENA Start Program creates specific challenges for the design
of the evaluation protocol. For example, the low-income families participating in the PHD Study
have unstable living arrangements or unforecastable work schedules. In such circumstances,
recruitment strategies face a trade-off. On the one hand, the longer the recruitment period, the
more likely we will locate an eligible family and consent the study’s participants. On the other
hand, the longer the recruitment period, the more likely parents who were recruited early would
report a change in the work schedule and could no longer attend group sessions when they had
agreed to participate in recruitment (to the LENA Start Program). Therefore, our team decided
to make an intense effort to recruit participants within three weeks and start the Program in the

5
Appendix B contains additional information about the PHD Study.

6
fourth week. In the three weeks of the recruitment effort, we managed to locate 94 (or 65%) of
the 145 parents in the control group and 95 (or 66%) of the 144 parents in the LENA Start
group’s invitation.

Next, we describe the protocol for the recruitment of participants. Consider a block with
twenty-six mothers, ordered in descending order of the draw from the uniform random variable.
Therefore, the mother ranked first is the one who drew the largest realization of the random
draw, while the mother with the smallest draw is ranked last. We invited the top thirteen mothers
to participate in the LENA Start Program. In addition, a few of the bottom thirteen may receive
an invitation to join if we need additional participants to form a LENA Start group.6

We invited the mothers at the bottom half of the ordered invitation list to participate in the
study’s control arm. First, we called the control mothers from bottom to top. Once invited, the
mothers assigned to the control group could accept or decline our invitation. The parents who
consented to participate in the evaluation study’s control arm agreed to participate in the study’s
data collection procedures, which we will explain in the next section.

We invited the mothers to participate in the study’s treatment arm at the top of the
ordered invitation list from top to bottom (i.e., the opposite for the control group). The mothers
could accept or decline our invitation. If they accepted the invitation, the research assistants
found a suitable schedule for the mothers to participate in the LENA Start sessions. If a mother
assigned to the treatment group rejected our invitation to join the LENA Start Program, our
research assistants invited that mother to participate as a control group member. We did so for
two reasons. First, we had little time to recruit parents for the evaluation study. Because of our
families’ unstable living arrangements, we knew it would not be feasible to contact a substantial
fraction of our eligible study participants within three weeks. This recruitment protocol helped
retain as many study participants as possible. Second, we expected a low demand for the
LENA Start Program. We formed our expectations from the literature that reports the demand
for parenting education programs (e.g., Kalil, 2014). Studies that follow the typical protocol do
not usually collect data from parents who refuse to participate as treatment group members. As
a result, we typically know very little about such parents. Because they join as members of the
control group (although no longer in a randomized fashion), we can learn more about these

6
We invited six control mothers to participate in the LENA Start Program to ensure we reached a
minimum number of mothers in some groups.

7
parents with our team’s protocol for this study. Also, we can still recover intent-to-treat treatment
effect parameters because the assignment variable ‫ݖ‬௜ is random.

As shown in Figure 1, once we found the parents, we tried to consent the parents to
participate in the study activities. As a result, nearly 75% (71 out of 94) of the parents in the
control group agreed to participate in the study. On the other hand, approximately 68% (65 out
of 95) of the treatment-group parents decided to participate in the study.

As we explained above, accepting the LENA Start Program was not random because
some families randomly assigned to receive the invitation declined the offer. Therefore, our
team invited a few control group members to have groups with ten families. As shown in Figure
1, the final acceptance list contained 61 families, 55 of which came from the study’s invitation
arm, and six came from the study’s control arm. However, only 39 of the 61 parents attended at
least one of the LENA Start Program sessions.

In Table 1, we show the results of the balance check. First, in Panel A, we compare the
PHD Study participants eligible for participation in the LENA Start Evaluation Study with non-
eligible ones. We find the differences between the control and invitation groups are small and
not statistically significant at the 5% level. However, we find that the counts of adult words are
lower for the children in the invitation group and that this difference has a p-value of 7.2%. We
note that the LENA measures indicate a lower quality of language environment because the
conversational turns are higher and exposure to TV is lower for the control group children (even
though the differences are not statistically significant). Second, we look at the post-attrition
sample in Panel B. There are no differences between the two groups except for the mother’s
marital status and HOME scores. In the post-attrition sample, those in the treatment group are
more likely to be single and unmarried. Appendix C provides more details about the balance
and attrition of our study sample.

B. Intervention

The LENA Start Program aims to improve parental linguistic input by improving parental
knowledge about the importance of the early language environment for language development
and providing tips for enriching the early language environment. 7 The Program lasts thirteen
weeks, and during this period, groups of ten to twenty families meet for about one hour a week

7
The website https://www.lena.org/lena-start/ provides an overview of the LENA Start Program.

8
to build family engagement and social capital. The program has five components: Education,
Coaching, Feedback, Book Reading, and Language Development Reporting.

Education: The program provides information about the importance of the language
environment for a child’s language development. Rowe and Goldin-Meadow (2009) show that
parents misperceive linguistic interaction with their young children as unimportant because
young children do not yet know how to verbalize. The program aims to affect these beliefs by
presenting parents with research documenting infants’ capacity to engage in social interaction
and that parents and infants routinely use nonverbal communication. For example, in one of the
sessions, the program presents the “Still Face Experiment” by Tronick et al. (1978). The
experiment involves an infant and a mother who remains expressionless for three minutes. The
finding from this experiment is that infants attempt many different forms of nonverbal
communication (facial or gesture expressions) to elicit reactions from their mothers. The infants
become upset with the mother’s lack of response and cry or start self-soothing behaviors. The
study remains one of the most replicated findings in developmental psychology, and it illustrates
to economically diverse parents how infants use nonverbal communication in social settings.

Coaching: Each session leverages “Talking Tips” to improve the child’s language
environment without requiring significant changes in the parents’ daily routines. Parents learn
these talking tips by watching video vignettes that show other parents interacting with an infant
or toddler. After observing other parents’ behavior, the parents reflect on how they could use
(with or without changes) these talking tips. As the program advances, the videos become more
complex, and parents assess and discuss, as a group, which strategies other parents employed
and whether parents in the vignettes missed any opportunities. Parents are encouraged to
share successful strategies when they return to a later session. They use their own experience
to generate new talking tips that other parents in the group can also practice. Coaching models
behavior and encourages parents to explore and identify new opportunities to improve the
child’s language environment.

The talking tips illustrate, in practice, vital aspects of a high-quality language

environment: joint (or shared) attention and speech recasting. The talking tips present easily
accessible information for families about how children learn language from what they hear most,
what interests them, conversations, and positive relationships. Joint attention is the shared
focus of two individuals on an object. An individual uses eye-gazing, pointing, or other verbal or
non-verbal communication to initiate a session of joint attention with another individual. Events
of joint attention promote language development because they provide a context that enables

9
children to associate meaning to a particular utterance, thus promoting word comprehension
and vocabulary expression (Bruner, 1983). For example, one of the tips is to “name things that
the child is interested in.” This tip reflects research that shows infants are more likely to engage
in joint attention when the parent refers to an object the child is playing with and far more
difficult when the object is outside the child’s attention (Rollins, 2003).

An event of speech recasting occurs when the adult repeats the child’s speech with
more detailed (or more correct) language. It allows the child to hear the accurate language, thus
helping the child figure out language syntax. The critical aspect of speech recasting is not to
force the child to repeat after the adult but rather for the adult to emphasize the linguistic feature
the child needs to learn (Cleave et al., 2015). Indeed, one of the tips is for parents to “repeat
and add to what they say and do.” In summary, the LENA Start Program builds its “Talking
Tips” component on the science that uncovers the contribution of joint attention and speech
recasting to language development.

Feedback: At the end of each Group Session, each parent receives a LENA recorder,
and each family completes a daylong recording with their children. When the parents return to
the next LENA Start session, they hand in the recorder to the LENA Start Coordinator, who
uploads the data and produces a feedback report. Parents have the chance to review their
information and notice the trends from one week to the next and the times of days of strength in
communication. The LENA Start Coordinator then offers to discuss the report with each parent.
The feedback follows a dialectical approach in which the Coordinators recognize that the
parents are doing the best they can and, at the same time, point out areas where the parents
can improve. Therefore, when discussing the report, the Coordinator first praises the parents for
their accomplishments (marked with stars). Second, the Coordinator identifies segments of the
day in which the language environment still has some growth potential.

To be able to produce this feedback, the report summarizes conversational turns

between adults and the focus child, adult words spoken around the focus child, which includes
speech other than the speech directed to the child (e.g., a parent talking on the telephone), and
the amount of time of exposure to electronics/TV, and self-reported reading minutes. The
feedback report presents this information aggregated to the entire day (for the last eight
recording days) and broken down hourly (for the previous recording day).

Figure 2 displays a report template. The report contains three rows and two columns.
The top row displays the percentiles (and counts) of adult words, the middle row shows the
percentiles (and counts) of conversational turns, and the bottom row presents data about

10
minutes of sounds from TV or other electronics. The left columns offer these statistics
aggregated by recording day for the last eight recording days. The right columns present the
statistics broken down by the hour for the previous recording day. The green text on the top
right corner of the feedback report contains a parent self-report number of reading minutes. In
addition, the feedback report includes the number of stars parents have received to date. The
parent gets a star when they meet pre-specified targets for adult words, conversational turns,
and reading minutes. The Coordinator praises the parent for the stars to recognize that they are
doing their best.

Next, the Coordinator uses the report to identify opportunities for improvement. For
example, when reviewing the information, the Coordinator will remark that the parent and the
child interacted a lot (as measured by conversational turns by the hour, as reported in the
middle row and right column of Figure 2) between 4 PM and 5 PM, but not so much at 11 AM
(or 1 PM). The fact that the information is objective, detailed, and actionable may provide
parents with opportunities to improve their child’s language environment. It also changes the
role of the Coordinator to be a support, not the one passing judgment on a parent’s skills. The
primary focus is on providing feedback about conversational turns. If there is much exposure to
TV/electronics, then the Coordinator will encourage parents to shift from TV to conversations
with the child.

Book reading: At each session, the parents receive children’s books. Each session
incorporates tips on reading books, talking about books to children, and encouraging parents to
read to their children. This part of the intervention aims to replicate the goals of programs such
as “Reach Out and Read” (Weitzman et al., 2004).

Reporting: Approximately once a month, the parents answer the Developmental

Snapshot (Gilkerson et al., 2018). The Snapshot is an instrument designed to evaluate
children’s language skills. It provides developmental age and percentile ranking information
compared to age-matched peers. The goal is for parents to start paying attention to their
children’s development and observing how their language skills have progressed. LENA Start
Coordinators review the Developmental Snapshot data with the parent and screen for signals of
severe language development delays that may require additional attention from an early
intervention specialist.

The parents who consented to participate in the LENA Start Program agreed to attend
the thirteen weekly sessions held at our Children’s Hospital of Philadelphia. A parent graduates
from the LENA Start Program if the parent satisfies two requirements. First, the parent attends

11
all of the first four sessions. Second, the parent had to complete at least five of the remaining
nine sessions.8

C. Measurement

The parents who consented to participate in the study agreed to provide data on their
children’s language environment and answer a brief survey questionnaire. The procedures and
instruments were identical for control and treatment groups and the same before (baseline) and
after the program (endline).

We describe the logistics of the recording activities carried out by families participating in
the study. The logistics described in this section relate to the data collection to evaluate the
LENA Start Program. As such, we describe a protocol applied to families in the study’s control
and treatment arms. Each parent received a LENA recorder and a child’s vest that had a front
pocket for the recorder. Parents recorded the child’s audio environment once at baseline and
once at the endline. We asked the parent to let the device run for at least 12 hours to record the
child’s audio environment. In addition, we asked parents to record the language environment on
a typical day for the child (e.g., not to record if the child was sick or an unusual event such as a
birthday party).

After the family had finished the recording session, we retrieved the recording device
and uploaded the audio file to a secure cloud server. The server contains software that
processes the audio data and automatically produces four different statistics: number of
conversational turns, number of child vocalizations, number of adult words, and minutes of
audio from the TV or other electronics.9 We focus our analysis on conversational turns and adult
words in audio segments with at least one conversational turn between the focus child and an
adult. We have standardized these variables to have mean zero and variance one in all of the
estimates we report below.

The LENA System identifies and labels individual adult and child utterances (also called
segments), among other sounds. One conversational turn is a speech-related vocal utterance
initiated by the focus child (or adult) with a response within five seconds. Vocal utterances may
include coos, squeals, babbles, and words. The number of conversational turns between the
adults in the household and the child is the most critical measure of the language environment

8
If the parent could not attend a session, they could contact the LENA Start Coordinator to schedule a
make-up session.
9
The estimates of the language environment variables LENA System produces are highly accurate (see
Gilkerson and Richards, 2008).

12
we use in our paper because it correlates with adult-child joint attention and speech recasting.10
Golinkoff et al. (2019) argue that “back-and-forth conversations that are both temporally and
topically contingent on the children’s contribution are the fuel that prime the learning of
language.”

The LENA System estimates the number of adult words spoken by adults based on
automated recognition of consonants and vowels and vocalization durations after filtering
unclear speech. However, unlike adult word statistics commonly used in established research
(e.g., Hart and Risley, 1995; Weisleder & Fernald, 2013), the LENA System estimate does not
distinguish between words directed to the child and words overheard by the child. Research
suggests that language learning from overheard speech does not occur until children are about
four or five years old (Messenger et al., 2015; Rollins, 2003).11 As explained by Golinkoff et al.
(2019), overheard speech does not support a rich early language environment because it
demands a lot more attention from young children (they have to stop doing what they are doing
and pay attention to what the other people are doing), it requires a level of social-cognitive skills
that children may not have developed yet (because they need to understand the intentions of
the adults involved in the conversation), and adult-directed speech is different from the child-
directed speech in content, tone of voice, and grammatical complexity. For this reason, our
analysis explores adult words in audio segments that have at least one conversational turn with
the focus child.

Finally, the LENA System identifies audio from the TV or other electronics as a separate
category of the child’s audio environment.

One of the innovations of this paper is that we have an indirect way of assessing
whether the improvements in conversational turns are “just for show” or reflect fundamental
changes in interactions between parents and children. This indirect way consists of using the
LENA Advanced-Data Extractor (ADEX) to break the audio file into segments that include
conversations between the focus child and an adult (male or female). In so doing, we can then
identify who initiated the conversation (the child or a male or a female adult) and the number of
conversational turns per segment as categorized by the segment initiator (the child, a female

10
See Goldin-Meadow et al. (2014); Harris et al. (2010); Malin et al. (2014); McGillion et al. (2017);
McGillion et al. (2017b); Reed et al. (2017); Romeo et al. (2018), and Tamis-LeMonda et al. (2014) for
additional insights on why conversational turns are so important for language development.
11
Studies of overheard speech may overestimate its impact on language development because they
typically eliminate or at least significantly reduce other stimuli in the lab. Hence, they minimize attentional
demands from children. See, for example, Yuan and Fisher (2009).

13
adult, or a male adult). Appendix D provides a detailed description of our procedures to clean
and create the final dataset for analysis.

We interviewed the parents who agreed to participate in the study to understand the
mechanisms of the intervention. The baseline and endline surveys occurred just before and
after the LENA Start Program, respectively. All 136 parents who consented to participate in the
study answered the survey questionnaire in the baseline, and 130 of the same parents did so in
the endline. The baseline and endline survey questionnaire had only eleven questions and
focused on measuring four potential mechanisms.

We measure four potential mechanisms. First, our survey questionnaire also measures
parental self-efficacy (Bandura, 1977), defined as the parent’s belief in their ability to perform
the parenting role competently. The scale has four statements, and we ask participants to
choose one (out of five) alternative that ranges from “least sure” to “very sure.” For example,
one item states, "I know what my child should be able to do at each age as they grow.” Coleman
and Karraker (1998) and Jones and Prinz (2005) summarize the literature and report that the
parental self-efficacy scores predict the child’s psychological functioning and adjustment. They
also found that parents with higher self-efficacy scores had higher parenting competence and
satisfaction levels (Leahy-Warren et al., 2012).

Second, we ask parents to report their social support perception because of the
intervention’s group nature. The scale has four items, and we ask parents to choose one (out of
five) alternative that describes the extent to which they agree with the statement. For example,
one statement is, “It is easy for me to talk with other parents about being a parent.” Previous
research has shown that positive social support from family and friends increases parenting
competence by providing encouragement and resources, particularly for first-time mothers
(Leahy-Warren et al., 2012).

We estimate an Item Response Theory Partial Credit Model (IRT-PCM) to produce self-
efficacy scores and sense of support scales. IRT-PCM is useful when researchers have ordinal
data. While allowing each item to have its discrimination parameter would be possible, we
chose this more parsimonious model because each scale has just four items. We provide
additional detail about these three measures in Appendix E.

Third, we measured parental knowledge of language development. We used the

instrument Suskind et al. (2016) developed to evaluate a parenting program’s impact on
parental linguistic input and knowledge. The questionnaire has 30 items divided into five

14
subscales. The first subscale covers parental perceptions about how children learn to talk.12 The
second subscale measures parental attitudes about reading to children.13 The third subscale
assesses parental perceptions about how children learn math.14 The fourth subscale evaluates
parental perceptions about the relationship between language development and school
readiness.15 The fifth and last subscale quantifies parental perceptions about TV and language
development.16 For each item, parents choose one alternative (among five) that describes the
extent to which they agree with the statement. The alternative ranges from “Strongly Disagree”
to “Strongly Agree.” We estimate the global and topic-specific scores.

Fourth, we adapted the procedure in Cunha et al. (2013) to elicit parental beliefs about
the importance of the language environment for language development. The elicitation
instrument has two items reflecting low and high scenarios of investments and asking mothers
to predict the child’s future language development. The first scenario describes a language
environment with many conversational turns between adults and children and little exposure to
TV. In contrast, the second scenario represents a low-quality language environment, with lots of
exposure to TV and little verbal exchange between adults and children. Finally, we ask parents
to predict the child’s language development at three years old. Parents choose one of the
following five alternatives: “Low,” “Low Average,” “Average,” “High Average,” and “High,” and we
assign these alternatives to percentiles 5, 25, 50, 75, and 95, respectively.

We follow Cunha et al. (2022) in constructing parental beliefs. First, we assume that
language development at age three follows a normal distribution with mean zero and variance
one. Then, we compute the scores of the standardized normal for each one of the percentiles.
Let ‫ݏ‬௜,ଵ and ‫ݏ‬௜,ଶ denote mother ݅’s score in the first and second scenarios, respectively. The
difference ο௜ ൌ ‫ݏ‬௜,ଵ െ ‫ݏ‬௜,ଶ represents our error-ridden measure of parental beliefs.

Cunha, Elo, and Culhane (2022) show that the estimated parental beliefs contain much
measurement error. Therefore, we combine the error-ridden measure above with three items

12
The second item in this subscale states that: “Children learn fewer words when adults talk with a warm
tone.”
13
The first item in this subscale states that “You cannot teach children anything new by reading them the
same book over and over.”
14
The third item in this subscale states that “Talking about the difference between tall and short teaches
toddlers about math.
15
The third item in this subscale states that “How many words a two-year-olds know can predict how well
they might do in kindergarten.”
16
The last item in this subscale states that “The more television children under two watch by themselves
the more words they learn.”

15
from the Suskind et al. (2016) knowledge scale.17 We factor analyze the four measures and
predict the factor score using the Bartlett formula. This factor score is our error-adjusted
estimate of parental beliefs.

D. Methods

In our impact analysis below, we estimate two different treatment effect parameters. The
first parameter is the “Intent-to-Treat” parameter (ITT), in which we use the binary variable ܼ
that captures random assignment to control (ܼ௜ ൌ 0) or invitation arm (ܼ௜ ൌ 1). Let ‫ܦ‬௜ ൌ 1 if the
family attended the LENA Start Program, and ‫ܦ‬௜ ൌ 0 otherwise. The second treatment effect
parameter is the “Local Average Treatment Effect (LATE),” which uses ܼ௜ as an instrumental
variable for ‫ܦ‬௜ .

Let ܻ௜,௧ denote child ݅’s conversational turns at period ‫ݐ‬. The variable ‫ܤ‬௜,௝ is equal to one
if mother ݅ is in randomization block ݆, and zero, otherwise. Let ܴ௜,௧ denote the vector that
contains information about child ݅’s LENA recording at period ‫ݐ‬, where ‫ ݐ‬ൌ 0 is the baseline, and
‫ ݐ‬ൌ 1 is the endline. This vector includes the recording duration and dummies for a recording
that occurred on Saturday or Sunday. Finally, the variable ܺ௜ captures demographic
characteristics. We estimate the following regression model to obtain the ITT:

ଵ ௃

ܻ௜,ଵ ൌ ߚ଴ ൅ ߚଵ ܼ௜ ൅ ߚଶ ܻ௜,଴ ൅ ܺ௜ ߚଷ ൅ ෍ ߚସା௧ ܴ௜,௧ ൅ ෍ ߛ௝ ‫ܤ‬௜,௝ ൅ ߳௜,ଵ . (1)

௧ୀ଴ ௝ୀଵ

We now turn to the analysis of the impacts of the LENA Start Program on parents who
attended at least one session. We estimate the LATE from the following two-stage least squares
regression model:

ଵ ௃

‫ܦ‬௜ ൌ ߙ଴ ൅ ߙଵ ܼ௜ ൅ ߙଶ ܻ௜,଴ ൅ ܺ௜ ߙଷ ൅ ෍ ߙସା௧ ܴ௜,௧ ൅ ෍ ߨ௝ ‫ܤ‬௜,௝ ൅ ߱௜,ଵ . (2)

௧ୀ଴ ௝ୀଵ

ଵ ௃

ܻ௜,ଵ ൌ ߚ଴ ൅ ߚଵ ‫ܦ‬௜ ൅ ߚଶ ܻ௜,଴ ൅ ܺ௜ ߚଷ ൅ ෍ ߚସା௧ ܴ௜,௧ ൅ ෍ ߛ௝ ‫ܤ‬௜,௝ ൅ ߳௜,ଵ . (3)

௧ୀ଴ ௝ୀଵ

17
We selected the three items with the strongest correlation with our error-ridden measure of parental
beliefs. They are items #22 (“Talking to children cannot make them smarter”), #23 (“How many words
three-year-olds know can predict how well they might do in kindergarten”), and #30 (“The more television
children under two watch by themselves the more words they learn”).

16
 For both ITT and LATE, we consider three variations of the models. The first model
imposes ߚଶ ൌ ߚଷ ൌ ߚସ ൌ 0. Thus, it does not include the lagged outcome, covariates for
demographic characteristics, or baseline recording information (Model 1). The subsequent
models have each one of these components, one at a time. In the second model, we include
demographic characteristics (so we relax the restriction that ߚଷ ൌ 0) (Model 2). Finally, we add
baseline conversational turn counts and recording information and estimate all the coefficients
in (1) (Model 3).

The error term ߳௜,ଵ is not necessarily independent across observations because of our
study design. For this reason, we cluster observations by recruitment groups for all of our
estimates of the standard errors. We believe that cluster is justified because programs will
attempt to form groups of parents whose children are of similar ages to encourage parents to
share experiences and exchange ideas about implementing the “Talking Tips.” However, due to
the insufficient number of clustered groups (we only have ten groups), we also provide
randomization inference p-values. We use the Young (2018) permutation test with 2,000
stratified clustered resamplings. We also report the conventional robust standard errors for all
the regression results for comparison purposes. We show the robustness of our point estimates
and the corresponding methods in Appendix F.

Let ‫ܯ‬௜,௞,௧ denote child ݅’s mediator ݇ at period ‫ݐ‬. We analyze two regression equations in
the mediation analysis. The first equation estimates the impact of the LENA Start Program on
each one of the mediators separately:
௃

‫ܯ‬௜,௞,ଵ ൌ ߜ௞,଴ ൅ ߜ௞,ଵ ܼ௜ ൅ ߜ௞,ଶ ‫ܯ‬௜,௞,଴ ൅ ܺ௜ ߜ௞,ଷ ൅ ෍ ߠ௞,௝ ‫ܩ‬௜,௝ ൅ ߟ௜,ଵ . (4)
௝ୀଵ

The parameter of interest is ߜ௞,ଵ in equation (4). The second equation expands equation (1) by
including mediator ݇:

ଵ ௃

ܻ௜,ଵ ൌ ߬௞,଴ ൅ ߬௞,ଵ ܼ௜ ൅ ߩ௞ ‫ܯ‬௜,௞,ଵ ൅ ߬௞,ଶ ܻ௜,଴ ൅ ܺ௜ ߬௞,ଷ ൅ ෍ ߬௞,௧ାସ ܴ௜,௧ ൅ ෍ ߶௞,௝ ‫ܩ‬௜,௝ ൅ ߞ௜,௞,ଵ . (5)
௧ୀ଴ ௝ୀଵ

In this equation, we are interested in the parameters ߬௞,ଵ and ߩ௞ . In this mediation exercise, the
parameter ߚଵ in (1) captures the LENA Start Program’s total effect on conversational turns. We
can decompose the total effect through the formula ߚଵ ൌ ߬௞,ଵ ൅ ߩ௞ ߜ௞,ଵ , where ߬௞,ଵ and ߩ௞ ߜ௞,ଵ are
the direct and indirect effects, respectively. As elsewhere in our analyses, we estimate and
present robust and clustered standard errors.

17
III. Results

A. Conversational Turns

In this section, we present the program’s impacts on conversational turns because it is

our primary variable of interest. Table 2 reports our findings regarding the impact of the LENA
Start Program on conversational turn counts. Panel A presents the OLS estimators of ߚଵ and ߚଶ
of variations of (1). According to the OLS estimator of Model 1, the LENA Start Program
increases conversational turns between adults and the focus child by over 13% of a standard
deviation, but it is not statistically significant. When we control for demographic characteristics,
the point estimate of the program’s intent-to-treat effect (ITT) increases to 21.5% of a standard
deviation, but it still is statistically insignificant. In model 3, when we control for baseline
conversational turns and recording characteristics, the ITT’s point estimate is 31.4% of a
standard deviation and statistically significant at a 5% confidence level.

Panel B in Table 2 presents the results for LATE. According to Model 1, the program’s
impact on conversational turns, estimated by the LATE parameter, is 23% of a standard
deviation. If we control for demographic characteristics, the LATE increases to 38.3% of a
standard deviation and is statistically significant at the 10% level. If we hold the conversational
turn counts at baseline constant, the LATE increases to 55.1% of a standard deviation and is
statistically significant at the 1% level.

These large impacts may arise because parents in the treatment group converse much
more with the children but only when the children wear a LENA recorder. It would be
challenging to design and implement an ethical study that comprehensively addresses this
concern as any such research would have to disclose all of the study procedures (including
recording the child’s language environment). We thus use the ADEX audio segmentation tool to
investigate whether children or adults initiate more conversations (audio segments). If the adults
change their behavior because of the recorder, then the differences in conversational turns
between the control and treatment groups will reflect the differences in conversations initiated
by the adults. On the other hand, if the adults apply what they have learned from the LENA Start
Program, they will let the child take the lead (which is one of the fourteen “Talking Tips”). In this
case, the changes in conversational turns arise because the children in the treatment group
initiate more conversations than their control counterparts, and the adults respond in ways that
require a response from the child. Therefore, the number of conversations (the number of
continuous audio segments with the focus child and an adult) and the number of conversational

18
turns increase because of the change in the parental responses to a conversation initiated by
the focus child. This behavior change is consistent with the LENA Start curriculum.

For this analysis, we use the lagged variable model described in equation (1) to estimate
the ITT parameter and the model described in equations (2) and (3) to compute the LATE
parameter. Panel A in Table 3 presents the results of our analysis of the initiation of segments
of conversation. The OLS and the 2SLS estimators show that the differences in conversational
turns arise because the LENA Start Program children initiate more conversations. According to
the ITT, the treatment group children begin nearly 40% of a standard deviation more
conversations than their control counterparts. The LATE parameters suggest that the impact is
69% of a standard deviation. In contrast, the estimates are statistically significant for
conversations initiated by female or male adults.

Panel B reports the number of conversational turns in segments initiated by the focus
child, a female adult, and a male adult. The ITT and the LATE parameters show that the
changes in conversations initiated by the focus child drive conversational turns. In addition, we
find suggestive evidence (for the LATE) that male adults increase the number of conversational
turns with children, even though there was only one male LENA Start Program participant. If
ever replicated, this finding suggests spillover within families.

In summary, the LENA Start parents do not initiate more conversations than the control
group’s parents. However, LENA Start children initiate more conversations, and when they do
so, the LENA Start parents respond in ways that lead to more conversational turns. These
findings reduce (but do not eliminate) the concern that the increase in conversational turns
occurs when they wear the device.

These results are significant because of the existing research on the differences
between high- and low-quality language environments. Hart and Risley (1995) document that
the disparity in words addressed to children results from children in higher-income households
initiating more conversations and adults responding in ways that lead to more conversational
turns and, as a result, more words directed to the child. This finding suggests that the language
environment is more affluent because parents talk about objects of interest to the child, and thus
the conversations occur during joint attention events.

B. Robustness

Appendix F extends our analyses to include three critical robustness checks. First, we
used lagged dependent variables in equations (1), (2), and (3) to account for the fact that the

19
families in the control group had higher counts of conversational turns at baseline (see Table
D2). Appendix F shows that the point estimates are similar if we use a fixed-effect specification.
For example, the ITT is .312 in the lagged dependent model but .323 in the fixed-effect
specification. However, the standard errors are larger, and the impacts are no longer significant
at the 5% level.

Second, we investigate how attrition impacted our findings (see Table F2). The ITT
reduces to .295 (from .312, a 5% reduction) if we use Inverse Probability Weight to reweigh
observations. However, it is .314 when we estimate a Heckman Selection model. The Lee
(2009) bounds indicate that the ITT is between .056 (with a standard error of .160) and .366
(with a standard error of .158). The lower bound is not statistically significant at the 5% level, but
the upper bound is.

Finally, recording duration enters linearly in equations (1) and (3). Therefore, we
considered specifications with polynomials of degrees two to four (see Table F3). When we do
so, we find that the ITT and LATE estimates are slightly larger then the linear specification we
report in Table 2.

C. Other Measures of the Quality of the Language Environment

The LENA System produces two additional measures of the quality of the language
environment: adult words and the amount of time of sounds from TV or other electronics.
Unfortunately, the raw measures are very noisy. For example, adult words may capture words in
a telephone conversation when the child is nearby. As we explain in Appendix A, this speech
contributes little to early language development.

We use the ADEX to reduce the noise in these measures (see Table G1). First, we
restrict adult words in audio segments where there is at least a conversational turn between the
child and an adult. When we do so, we find that the ITT is .263 but not statistically significant,
while the LATE is .468 and significant at the 5% level.

Second, we restrict adult words in audio segments initiated by the focus child. These
segments drive the impact on conversational turns (see Table 3). When we do so, we find that
the ITT and LATE are .495 and .886, respectively. Also, they are significant at the 5% and 1%
levels, respectively. This evidence reinforces the findings on conversational turns.

In contrast, we do not find any impact of the program on TV or other electronics. The
point estimates are small (see Table G2).

20
D. Mediation Analysis

Finally, we study the potential mechanisms of the impacts of the program. As described
in Section II.C., we measured four possible mechanisms parental self-efficacy, parental sense
of social support, parental knowledge, and parental beliefs. Appendix C contains detailed
information about how we operationalize these constructs.

Table 4 presents our estimates for the parameter ߜ௞,ଵ in equation (4). The ITTs for
parental self-efficacy and parental sense of social support are close to zero and not statistically
significant. In contrast, the ITT for parental knowledge is .293, and borderline statistically
significant. The ITTs for parental beliefs are .306 and .451 for the error-ridden and factor score
measures, respectively, and are statistically significant at the 10% and 5% levels, respectively.
The LATE parameters also rule out any impacts of the LENA Start Program on parental self-
efficacy or sense of social support. In addition, it confirms that the program impacted parental
knowledge and parental beliefs.

Table 5 presents our estimates of the parameters ߩ௞ in equation (5). The ITT reduces by
20% when we use the error-ridden measure of parental beliefs, but it is still statistically
significant, while the parameter ߩ௞ is 0.098 and not statistically significant. However, the ITT
reduces by nearly half when we add the factor score for parental beliefs, which is no longer
significant. In this case, ߩ௞ is 0.198 and significant at the 5% level.

We execute a Monte Carlo simulation to estimate the distribution of the indirect effect
ߩ௞ ߜ௞,ଵ . For this analysis, we focus on the factor score as the relevant measure. Figure 3
presents our findings for the models in which we estimate clustered (top) or robust (bottom)
standard errors. We estimate that the indirect effect is 13.7% of the total effect. The 95%
confidence intervals are [0.026, 0.288] and [0.002, 0.321] for the models with clustered and
robust standard errors. Thus, we conclude that parental beliefs partially mediate the effects of
the intervention.

In contrast, table G3 presents the results for the other three mediators. We cannot reject
the null that the coefficient ߩ௞ is zero, which suggests that the other three variables do not
mediate the intervention’s impact on the quality of the language environment.

IV. Conclusion

The LENA Start Program improves children’s language environment. These findings
matter because parental linguistic input correlates with early language processing skills,
vocabulary growth, higher language development in early adolescence, and more robust

21
activation of Broca’s brain area, which, in turn, are linked to literacy development. In addition,
differences in parental response to conversations initiated by the child drive the program’s
impacts.

We tested four different mechanisms. First, we found no evidence in favor of parental

self-efficacy, parental social support, or parental knowledge. Instead, we found evidence that
the program’s impact is due to improvements in parental beliefs about the importance of a rich
early language environment for a child’s language development. Our findings indicate that
parental beliefs are malleable and that changes in parental beliefs can change investment
behaviors. This result is consistent with the findings of the evaluation of the Nadia Es Perfecto –
Intensive (NES-I, Carneiro et al., 2019) and the 3T Home Visitation Program (List et al., 2021).
These interventions feature unstructured curricula that include education, coaching, and
feedback.

In contrast, Attanasio et al. (2019) concluded that Colombia’s Reach Up Program did not
change beliefs. There are at least three explanations for this difference: Curriculum, target
population, and implementation fidelity. In common, the Reach Up and LENA Start interventions
provide education and coaching. In contrast, Reach Up contains a highly structured curriculum
with much less opportunity for feedback. Therefore, one possible explanation is that lack of
structure or objective feedback is crucial for impacting parental beliefs.

Another possible explanation is the target population. For example, research shows that
the original implementation of the Jamaica Home Visiting Program led to significant short-term
effects on language and cognition (Grantham-McGregor et al., 1991) and long-term effects on
relevant socio-economic outcomes (Gertler et al., 2014; Gertler et al., 2021). Attanasio et al.
(2022) explain that the intervention in Jamaica targeted undernourished or severely
undernourished children. Interestingly, in Colombia, the program targeted low-income families
who participated in a Conditional Cash Transfer program, and there was no indication that the
children in these families were undernourished. This explanation suggests that the Reach Up
Program did not change parental beliefs because of targeting. 18

Finally, we must emphasize that the implementation of the Jamaica Home-Visiting

Program, the 3T Home Visiting Intervention, and the LENA Start Program occurred in research
settings that are human-capital abundant. In contrast, the Reach Up implementation in
Colombia occurred within a real-world setting and explored human resources available to local

18
See also Sylvia et al. (2021).

22
communities. For example, the home visitors were mothers who resided in the same
community. However, we remark that the Reach Up implementation in Bangladesh (Hamadani
et al., 2006) and China (Heckman et al., 2020) also used resources available in the community.
While the latter two studies did not measure parental beliefs, they produced impacts on
language and cognition much greater than Colombia’s and aligned with the Jamaica Home
Visiting Program’s impacts. Carneiro et al. (2019) evaluated the NEP-I program implemented
with community resources and found positive impacts on beliefs. Nonetheless, these authors
document that the Chilean government spent resources recruiting and training the NEP-I
workforce.

These contrasting findings suggest another explanation. The Reach Up Program did not
change parental beliefs because it was implemented with low fidelity. Unfortunately, we cannot
rule out this explanation because no data measures the quality of these programs when they
occur in real-world settings. However, such data is crucial to advance the science of the
determinants of early investments and the formulation of public policy that fosters human capital
formation.

We conclude by suggesting two avenues of future research. First, this paper has
established that the LENA Start Program impacted the language environment in the short run.
Future research should investigate if this impact persists and, if so, whether it translates into
improved language comprehension and literacy skills in the long run. Such research would
inform the debate about the long-term impacts of early interventions (e.g., Bailey et al., 2020;
García & Heckman, 2022).

Second, our LENA Start Program’s implementation did not occur in a real-world setting.
Thus, future research should investigate the LENA Start Program’s implementation fidelity in a
typical setting (e.g., as implemented by a school district or public library). If the execution is high
quality, researchers should measure the impacts on parental beliefs and language outcomes in
the short and long run. This research is vital to shedding light on vital scientific questions about
the malleability of parental beliefs and the formulation of public policy to foster early human
capital formation.

23
References
Andrew, Alison, Orazio Attanasio, Emla Fitzsimons, Sally M. Grantham-McGregor, Costas
Meghir, and Marta Rubio-Codina. "Impacts 2 Years after a Scalable Early Childhood
Development Intervention to Increase Psychosocial Stimulation in the Home: A Follow-
up of a Cluster Randomised Controlled Trial in Colombia." PLoS Medicine 15, no. 4
(2018).
Attanasio, Orazio, Flavio Cunha, and Pamela Jervis. "Subjective Parental Beliefs: Their
Measurement and Role." In NBER Working Paper 26516, 2019.
Attanasio, Orazio P., Camila Fernandes, Emla O. A. Fitzsimons, Sally M. Grantham-McGregor,
Costas Meghir, and Marta Rubio-Codina. "Using the Infrastructure of a Conditional Cash
Transfer Program to Deliver a Scalable Integrated Early Child Development Program in
Colombia: Cluster Randomized Controlled Trial." British Medical Journal 349, no. g5785
(2014).
Bailey, Drew, Greg Duncan, Flavio Cunha, Barbara Foorman, and David Yeager. "Persistence
and Fadeout of Educational Intervention Effects: Mechanisms and Potential Solutions."
Psychological Science in the Public Interest 21, no. 2 (2020): 55-97.
Bandura, A. "Self-Efficacy - toward a Unifying Theory of Behavioral Change." Psychological
Review 84, no. 2 (1977): 191-215.
Bhalotra, Sonia, Adeline Delavande, Paulino Font Gilabert, and Joanna Maselko. "Maternal
Investments in Children: The Role of Expected Effort and Returns." In IZA Institute of
Labor Economics Discussion Paper 13056, 2020.
Biroli, Pietro, Teodora Boneva, Akash Raja, and Chris Rauh. "Parental Beliefs About Returns to
Child Health Investments." Journal of Econometrics 231, no. 1 (2022): 33-57.
Boneva, Teodora, and Chris Rauh. "Parental Beliefs About Returns to Educational
Investments—the Later the Better?" Journal of the European Economic Association, 16,
no. 6 (2018): 1669-711.
Bradley, Robert H., and Bettye M. Caldwell. "The Relation of Home-Environment, Cognitive
Competence, and Iq among Males and Females." Child Development 51, no. 4 (1980):
1140-48.
Bruner, Jerome. Child's Talk: Learning to Use Language. . New York, NY: W.W. Norton & Co.,
1983.
Carneiro, Pedro, Emanuela Gallaso, Italo L. García, Paula Bedregal, and Miguel Cordero.
"Parental Beliefs, Investments, and Child Development: Evidence from a Large-Scale

24
 Experiment." In Discussion Paper Series, edited by IZA Institute of Labor Economics,
2019.
Cleave, Patricia L., Stephanie D. Becker, Maura K. Curran, Amanda J O Van Horn, and Marc
E Fey. "The Efficacy of Recasts in Language Intervention: A Systematic Review and
Meta-Analysis " American Journal of Speech-Language Pathology 24 (2015): 237-55.
Coleman, Priscilla K., and Katherine H. Karraker. "Self-Efficacy and Parenting Quality: Findings
and Future Applications." Developmental Review 18, no. 1 (1998): 47-85.
Cunha, Flavio, Irma Elo, and Jennifer Culhane. "Eliciting Maternal Expectations About the
Technology of Cognitive Skill Formation." In NBER Working Paper 19144, 2013.
———. "Maternal Subjective Expectations About the Technology of Skill Formation Predict
Investments in Children One Year Later." Journal of Econometrics 231, no. 1 (2022): 3-
32.
Duncan, Greg, and Richard J. Murnane. Whither Opportunity?: Rising Inequality, Schools, and
Children's Life Chances (Copublished with the Spencer Foundation) New York, NY.:
Russell Sage Foundation, 2011.
Fernald, Anne, Virginia A. Marchman, and Adriana Weisleder. "Ses Differences in Language
Processing Skill and Vocabulary Are Evident at 18 Months." Developmental Science 16,
no. 2 (2012): 234-48.
Foorman, Barbara R. , Yaacov Petscher, and Sarah Herrera. "Unique and Common Effects of
Decoding and Language Factors in Predicting Reading Comprehension in Grades 1–
10." Learning and Individual Differences 63, no. 1 (2018): 12-23.
García, Jorge Luis, and James J. Heckman. "Parenting Promotes Social Mobility within and
across Generations." In NBER Working Paper, 2022.
Gertler, P., J. Heckman, R. Pinto, A. Zanolini, C. Vermeersch, S. Walker, S. M. Chang, and S.
Grantham-McGregor. "Labor Market Returns to an Early Childhood Stimulation
Intervention in Jamaica." Science 344, no. 6187 (2014): 998-1001.
Gertler, Paul, James J Heckman, Rodrigo Pinto, Susan M Chang, Sally M Grantham-
McGregor, Christel Vermeersch, Susan Walker, and Amika Wright. "Effect of the
Jamaica Early Childhood Stimulation Intervention on Labor Market Outcomes at Age
31." In NBER Working Paper, 2021.
Gilkerson, Jill, and Jeffrey A. Richards. "The Lena Natural Language Study." 2008.
Gilkerson, Jill, Jeffrey A. Richards, Steven F. Warren, D. Kimbrough Oller, Rosemary Russo,
and Betty Vohr. "Language Experience in the Second Year of Life and Language
Outcomes in Late Childhood." Pediatrics 142, no. 4 (2018).

25
Golinkoff, Roberta M., Erika Hoff, Meredith L. Rowe, Catherine S. Tamis-LeMonda, and Kathy
Hirsh-Pasek. "Language Matters: Denying the Existence of the 30-Million-Word Gap Has
Serious Consequences." Child Development 0, no. 0 (2019).
Gough, Philip B., and William E. Tunmer. "Decoding, Reading, and Reading Disability."
Remedial and Special Education 7, no. 1 (1986): 6-10.
Grantham-McGregor, Sally, Christine A. Powell, Susan P. Walker, and John H. Himes.
"Nutritional Supplementation, Psychosocial Stimulation, and Mental Development of
Stunted Children: The Jamaican Study." The Lancet 338, no. 8758 (1991): 1-5.
Hamadani, Jena D., Syed N. Huda, Fahmida Khatun, and Sally M. Grantham-McGregor.
"Psychosocial Stimulation Improves the Development of Undernourished Children in
Rural Bangladesh." Journal of Nutrition 136, no. 10 (2006): 2645-52.
Hanushek, Erik A., Guido Schwerdt, Simon Wiederhold, and Ludger Woessmann. "Returns to
Skills around the World: Evidence from Piaac." European Economic Review 73 (2015):
103-30.
Hart, Betty, and Todd Risley. Meaningful Differences in the Everyday Experience of Young
American Children. Baltimore, MD: Paul H. Brookes Publishing Company, 1995.
Heckman, James J. , Margaret L. Holland, Kevin K. Makino, Rodrigo Pinto, and Maria
Rosales-Rueda. "An Analysis of the Memphis Nurse-Family Partnership Program." In
NBER Working Paper 23610, 2017.
Heckman, James J. , Bei Liu, Mai Lu, and Jin Zhou. "The Impacts of a Prototypical Home
Visiting Program on Child Skills." In NBER Working Paper 27356, 2020.
Hoff, Erika. "How Social Contexts Support and Shape Language Development." Developmental
Review 26, no. 1 (2006): 55-88.
———. "The Specificity of Environmental Influence: Socioeconomic Status Affects Early
Vocabulary Development Via Maternal Speech." Child Development 74, no. 5 (2003):
1368-78.
Huttenlocher, Janellen, Marina Vasilyeva, Heidi R. Waterfall, Jack L. Vevea, and Larry V.
Hedges. "The Varieties of Speech to Young Children." Developmental Psychology 43,
no. 5 (2007): 1062-83.
Huttenlocher, Janellen, Heidi R. Waterfall, Marina Vasilyeva, Jack L. Vevea, and Larry V.
Hedges. "Sources of Variability in Children's Language Growth." Cognitive Psychology
61, no. 4 (2010): 343-65.
Jones, Tracy L., and Ronald J. Prinz. "Potential Roles of Parental Self-Efficacy in Parent and
Child Adjustment: A Review." Clinical Psychology Review 25, no. 3 (2005): 341-63.

26
Kalil, Ariel "Inequality Begins at Home: The Role of Parenting in the Diverging Destinies of Rich
and Poor Children." In Families in an Era of Increasing Inequality: Diverging Destinies,
edited by Paul R. Amato, Alan Booth, Susan M. McHale and Jennifer Van Hook, 63-82.
New York: Springer, 2014.
Leahy-Warren, Patricia, Geraldine McCarthy, and Paul Corcoran. "First-Time Mothers: Social
Support, Maternal Parental Self-Efficacy, and Postnatal Depression." Journal of Clinical
Nursing 21, no. 3-4 (2012): 388-97.
Lee, David S. . "Training, Wages, and Sample Selection: Estimating Sharp Bounds on
Treatment Effects." Review of Economic Studies 76, no. 3 (2009): 1071-102.
Leech, Kathryn A., Ran Wei, Jeffrey R Harring, and Meredith L Rowe. "A Brief Parent-
Focused Intervention to Improve Preschoolers' Conversational Skills and School
Readiness." Developmental Psychology 54, no. 1 (2018): 15-28.
List, John A., Julie Pernaudet, and Dana L. Suskind. "It All Starts with Beliefs: Addressing the
Roots of Educational Inequities by Changing Parental Beliefs." edited by NBER Working
Paper, 2021.
McGillion, Michelle , Julian M Pine, Jane S Herbert, and Danielle Matthews. "A Randomised
Controlled Trial to Test the Effect of Promoting Caregiver Contingent Talk on Language
Development in Infants from Diverse Socioeconomic Status Backgrounds." Journal of
child psychology and psychiatry, and allied disciplines 58, no. 10 (2017): 1122-31.
Messenger, K., S. Yuan, and C. Fisher. "Learning Verb Syntax Via Listening: New Evidence
from 22-Month-Olds." Language Learning and Development 11, no. 4 (2015): 356-68.
Murnane, Richard J., John B. Willett, Yves Duhaldeborde, and John H. Tyler. "How Important
Are the Cognitive Skills of Teenagers in Predicting Subsequent Earnings?" Journal of
Policy Analysis and Management 19, no. 4 (2000): 547-68.
OECD. "Skills Matter: Further Results from the Survey of Adult Skills." In OECD Skills Studies.
Paris, 2016.
Olds, David L, Harriet Kitzman, Robert Cole, JoAnn Robinson, Kimberly Sidora, Dennis W
Luckey, Charles R Henderson Jr, Carole Hanks, Jessica Bondy, and John Holmberg.
"Effects of Nurse Home-Visiting on Maternal Life Course and Child Development: Age 6
Follow-up Results of a Randomized Trial." Pediatrics 114, no. 6 (2004): 1550-59.
Pan, Barbara A., Meredith L. Rowe, Judith D. Singer, and Catherine E. Snow. "Maternal
Correlates of Growth in Toddler Vocabulary Production in Low-Income Families." Child
Development 76, no. 4 (2005): 763-82.

27
Ramirez, Naja Ferjan, Sarah Roseberry Lytle, and Patricia K. Kuhl. "Parent Coaching Increases
Conversational Turns and Advances Infant Language Development." Proceedings of the
National Academy of Sciences of the United States of America 117, no. 7 (2020): 3484-
91.
Rollins, Pamela R. "Caregivers' Contingent Comments to 9-Month-Old Infants: Relationships
with Later Language." Applied Psycholinguistics 24, no. 2 (2003): 221-34.
Romeo, R. R., J. A. Leonard, S. T. Robinson, M. R. West, A. P. Mackey, M. L. Rowe, and J. D.
E. Gabrieli. "Beyond the 30-Million-Word Gap: Children's Conversational Exposure Is
Associated with Language-Related Brain Function." Psychological Science 29, no. 5
(2018): 700-10.
Rowe, Meredith L. , and Kathryn A. Leech. "A Parent Intervention with a Growth Mindset
Approach Improves Children's Early Gesture and Vocabulary Development."
Developmental science 22, no. 4 (2019): e12792.
Rowe, Meredith L., and Susan Goldin-Meadow. "Early Gesture Selectively Predicts Later
Language Learning." Developmental Science 12, no. 1 (2009): 182-87.
Scarborough, Hollis S, Susan Neuman, and David Dickinson. "Connecting Early Language and
Literacy to Later Reading (Dis) Abilities: Evidence, Theory, and Practice." Approaching
difficulties in literacy development: Assessment, pedagogy and programmes 10 (2009):
23-38.
Suskind, Dana L. , Kristin R. Leffel, Eileen Graf, Marc W. Hernandez, Elizabeth A. Gunderson,
Shannon G. Sapolich, Elizabeth Suskind, Lindsey Leininger, Susan Goldin-Meadow, and
Susan C. Levine. "A Parent-Directed Language Intervention for Children of Low
Socioeconomic Status: A Randomized Controlled Pilot Study." Journal of Child
Language 43, no. 2 (2016): 366-406.
Suskind, Dana L., Kristin R. Leffel, Marc W. Hernandez, Shannon G. Sapolich, Elizabeth
Suskind, Erin Kirkham, and Patric Meehan. "An Exploratory Study of “Quantitative
Linguistic Feedback”: Effect of Lena Feedback on Adult Language Production."
Communication Disorders Quarterly 34, no. 4 (2013): 199-209.
Sylvia, Sean , Nele Warrinnier, Renfu Luo, Ai Yue, Orazio Attanasio, Alexis Medina, and
Scott Rozelle. "From Quantity to Quality: Delivering a Home-Based Parenting
Intervention through China’s Family Planning Cadres." The Economic Journal 131, no.
635 (2021): 1365-400.

28
Tronick, Edward , Heidelise Als, Lauren Adamson, Susan. Wise, and T. Berry Brazelton. "The
Infant's Response to Entrapment between Contradictory Messages in Face-to-Face
Interaction." Child and Adolescent Psychiatry 17, no. 1 (1978): 1-13.
Vignoles, Anna, Augustin De Coulon, and Oscar Marcenaro-Gutierrez. "The Value of Basic
Skills in the British Labour Market." Oxford Economic Papers-New Series 63, no. 1
(2011): 27-48.
Weisleder, A., and A. Fernald. "Talking to Children Matters: Early Language Experience
Strengthens Processing and Builds Vocabulary." Psychological Science 24, no. 11
(2013): 2143-52.
Weitzman, Carol Cohen, Lisa Roy, Theodore Walls, and Ricarda Tomlin. "More Evidence for
Reach out and Read: A Home-Based Study." Pediatrics 113, no. 5 (2004): 1248-53.
Wooldridge, Jeffrey M. Econometric Analysis of Cross Section and Panel Data: MIT press,
2010.
Young, Alwyn. "Channeling Fisher: Randomization Tests and the Statistical Insignificance of
Seemingly Significant Experimental Results." The Quarterly Journal of Economics 134,
no. 2 (2018): 557-98.
Zhang, Yiwen, Xiaojuan Xu, Fan Jiang, Jill Gilkerson, Dongxin Xu, Jeffrey A. Richards, James
Hamberger, and Keith Topping. "Effects of Quantitative Linguistic Feedback to
Caregivers of Young Children: A Pilot Study in China." Communication Disorders
Quarterly 37, no. 1 (2015): 16-24.

29
Figure 1. — LEME Study: Eligibility, Random Assignment, Recruitment, Consent, Final
Assignment, Graduation.
Figure 2. — LENA Start Objective Feedback Report.
Figure 3. — Distribution of the Indirect Effect of the factor score of parental beliefs, clustered
(top) and robust (bottom) standard errors.
 TABLE 1
Balance Test Results of the LENA Start Program
Panel A: Baseline sample
Invitation to LENA
Control p-value
Start
(1) (2) (1) vs (2)
VARIABLES N Mean N Mean Robust Clustered
Household income is less than or 142 0.704 145 0.752 0.368 0.345
equal to twice the FPL1 (0.458) (0.434)
Mother has attended some higher 144 0.097 145 0.103 0.861 0.906
education program (0.297) (0.306)
144 0.153 145 0.200 0.294 0.354
Mother is Hispanic
(0.361) (0.401)
144 0.625 145 0.579 0.429 0.477
Mother is Non-Hispanic black
(0.486) (0.495)
144 0.139 145 0.159 0.639 0.689
Mother is Non-Hispanic white
(0.347) (0.367)
144 0.729 145 0.655 0.174 0.263
Mother is single
(0.446) (0.477)
144 0.063 145 0.083 0.509 0.454
Mother is cohabiting
(0.243) (0.276)
144 0.208 145 0.262 0.283 0.373
Mother is married
(0.408) (0.441)
133 -0.204 132 -0.044 0.265 0.203
Standardized HOME Score
(1.383) (0.887)
47 283.279 54 317.517 0.292 0.095
Conversational Turns at 9 months
(191.053) (120.901)
47 13761.77 54 16260.48 0.072 0.037
Adult Word Counts at 9 months
(6981.438) (6804.080)
47 7308.39 54 7052.62 0.795 0.727
Seconds of Exposure to TV
(5007.889) (4813.342)
Standardized Language Score from 137 -0.443 134 -0.502 0.521 0.469
BSID2 (0.708) (0.816)
Child currently being cared for by 131 0.718 132 0.659 0.308 0.398
someone else regularly (0.452) (0.476)
Number of different regular childcare 131 1.092 132 0.985 0.360 0.401
arrangements (0.932) (0.957)
 TABLE 1
Balance Test Results of the LENA Start Program
Panel B: Post-attrition sample
Invitation to LENA
Control p-value
Start
(3) (4) (3) vs (4)
VARIABLES N Mean N Mean Robust Clustered
Household income is less than or 63 0.825 71 0.775 0.466 0.428
equal to twice the FPL1 (0.383) (0.421)
Mother has attended some higher 65 0.031 71 0.070 0.291 0.404
education program (0.174) (0.258)
65 0.138 71 0.183 0.482 0.394
Mother is Hispanic
0.348 (0.390)
65 0.631 71 0.606 0.765 0.813
Mother is Non-Hispanic black
(0.486) (0.492)
65 0.123 71 0.155 0.594 0.542
Mother is Non-Hispanic white
(0.331) (0.364)
65 0.831 71 0.676 0.036 0.026
Mother is single
(0.378) (0.471)
65 0.062 71 0.070 0.836 0.828
Mother is cohabiting
(0.242) (0.258)
65 0.108 71 0.254 0.026 0.032
Mother is married
(0.312) (0.438)
61 -0.351 68 -0.015 0.097 0.067
Standardized HOME Score
(1.429) (0.686)
25 278.407 36 327.929 0.295 0.210
Conversational Turns at 9 months
(214.376) (115.985)
25 14958.98 36 17536.44 0.195 0.132
Adult Word Counts at 9 months
(7966.382) (6942.454)
25 8272.94 36 6798.43 0.275 0.204
Seconds of Exposure to TV
(5481.453) (4631.373)
Standardized Language Score from 64 -0.484 71 -0.505 0.872 0.881
BSID2 (0.667) (0.789)
Child currently being cared for by 60 0.700 67 0.672 0.733 0.762
someone else regularly (0.462) (0.473)
Number of different regular childcare 60 1.100 67 1.015 0.637 0.726
arrangements (1.053) (0.961)
NOTE. — Standard deviation in parenthesis. The p-values are obtained from item-wise regressions
using the variables labeled in the first column as the dependent variables and the random assignment
indicator as the only independent variable. We report the p-values for both robust and clustered
uncertainty estimates clustered at the recruiment group level.
1
Federal Poverty Line.
2
Bayley Scale of Infant Development.
 TABLE 2
Impact of the LENA Start Program on Conversational Turns
Dependent Variable: Conversational Turns
Variables Model 1 Model 2 Model 3
Panel A: OLS Estimator
Random assignment to control or invitation arm 0.130 0.215 0.314**
Clustered SE (0.187) (0.160) (0.113)
Robust SE (0.160) (0.163) (0.152)
Randomization inference p-value 0.462 0.257 0.072
Observations 104 102 90
R-squared 0.452 0.497 0.681
Panel B: 2SLS Estimator
Attendance to the LENA Start program 0.232 0.383* 0.551***
Clustered SE (0.286) (0.211) (0.125)
Robust SE (0.268) (0.264) (0.233)
Observations 104 102 90
R-squared 0.445 0.479 0.680
Recording duration at endline Y Y Y
Dummies for Saturday and Sunday at endline Y Y Y
Demographic characteristics N Y Y
Conversational Turn Counts at baseline N N Y
Recording duration at baseline N N Y
Dummies for Saturday and Sunday at baseline N N Y

NOTE. — Standard errors in parentheses. We report clustered and robust standard errors, and
we report the most conservative significance level. Clustered standard errors are clustered at
the recruiment group level. We also report the randomization inference p-value after 2,000
stratified clustered resampling. Panel A reports the estimates obtained from ordinary least
squares (OLS) regressions. Panel B reports the estimates obtained from two-stage least
squares regression (2SLS) regressions using treatment assignment indicator as an instrument
for attendance. We add the following variables to control for differences in recording sessions.
First, we control for recording duration in endline, a dummy variable for a recording done on
Saturday, and another dummy variable for a recording done on Sunday (Model 1) . Second, we
additionally add demographic characteristics (Model 2) including a dummy for maternal year of
birth between 1978 and 1987; dummy for Hispanic ethnicity; dummy for non-Hispanic black;
dummy for single mother; dummy for cohabiting mother; dummy for income below two times
the federal poverty line; dummy for mothers with some postsecondary education; dummy for
male child; age of the child at the date of the recording; and dummy for randomization group
(block). Third, we additionally add conversational turn counts at baseline, recoding duration at
baseline, and dummies for Saturday and Sunday at baseline (Model 3).

* Significant at the 10% level.

** Significant at the 5% level.
*** Significant at the 1% level.
 TABLE 3
Impact of the LENA Start Program on Initiation of Conversations in Audio Segments with Focus Child
and an Adult
Panel A: Number of Audio Segments
Number of Number of Number of
Segments Segments Segments
Initiated by the Initiated by a Initiated by a
Focus Child Female Adult Male Adult
OLS Estimator
Random assignment to control or invitation arm 0.394** 0.118 0.248
Clustered SE (0.155) (0.178) (0.177)
Robust SE (0.162) (0.153) (0.169)
Randomization inference p-value 0.012 0.479 0.174
Observations 87 87 87
R-squared 0.749 0.692 0.679
2SLS Estimator
Attendance to the LENA Start program 0.690*** 0.203 0.432*
Clustered SE (0.172) (0.228) (0.234)
Robust SE (0.230) (0.215) (0.245)
Observations 87 87 87
R-squared 0.745 0.706 0.675
 TABLE 3
Impact of the LENA Start Program on Initiation of Conversations in Audio Segments with Focus Child
and an Adult
Panel B: Number of Conversational Turns by Audio
Segment Initiator
Conversational Conversational Conversational
Turns in Turns in Turns in
Segments Segments Segments
Initiated by the Initiated by a Initiated by a
Focus Child Female Adult Male Adult
OLS Estimator
Random assignment to control or invitation arm 0.488** 0.238 0.331
Clustered SE (0.191) (0.217) (0.202)
Robust SE (0.169) (0.167) (0.181)
Randomization inference p-value 0.002 0.214 0.111
Observations 87 87 87
R-squared 0.728 0.563 0.624
2SLS Estimator
Attendance to the LENA Start program 0.862*** 0.414 0.572**
Clustered SE (0.209) (0.256) (0.223)
Robust SE (0.245) (0.238) (0.256)
Observations 87 87 87
R-squared 0.726 0.592 0.637

NOTE. — Standard errors in parentheses. We report clustered and robust standard errors, and we
report the most conservative significance level. Clustered standard errors are clustered at the
recruiment group level. We also report the randomization inference p-value after 2,000 stratified
clustered resampling. Panel A reports the effects on the number of audio segments. Panel B reports
the effects on the number of conversational turns by audio segment initiator. For both outcome
measures, we report estimates obtained from ordinary least squares (OLS) regressions and two-stage
least squares regression (2SLS) regressions using the treatment assignment indicator as an instrument
for attendance. In all the regressions, we add the following variables to control for differences in
recording sessions. We add the lagged dependent variable and we control for recording duration in
baseline and endline. We add a dummy variable for a recording done on Saturday and another dummy
variable for a recording done on Sunday, both for baseline and endline. Additionally, we add a dummy
for maternal year of birth between 1978 and 1987; dummy for Hispanic ethnicity; dummy for non-
Hispanic black; dummy for single mother; dummy for cohabiting mother; dummy for income below two
times the federal poverty line; dummy for mothers with some postsecondary education; dummy for male
child; age of the child at the date of the recording; and dummy for randomization group (block).

* Significant at the 10% level.

** Significant at the 5% level.
*** Significant at the 1% level.
 TABLE 4
Investigation of Potential Mechanisms of the LENA Start Program
Parental Self- Parental Sense Parental Parental Beliefs
Efficacy of Social Knowledge
Support Error-Ridden Factor Score
VARIABLES
Panel A: OLS Estimator
Random assignment to control or invitation arm -0.030 -0.077 0.293 0.306* 0.451**
Clustered SE (0.158) (0.153) (0.185) (0.151) (0.176)
Robust SE (0.131) (0.176) (0.139) (0.154) (0.143)
Randomization inference p-value 0.834 0.665 0.037 0.069 0.003
Observations 128 128 128 128 128
R-squared 0.569 0.352 0.614 0.318 0.377
Panel B: 2SLS Estimator
Attendance to the LENA Start program -0.057 -0.146 0.564** 0.586** 0.859***
Clustered SE (0.258) (0.240) (0.264) (0.246) (0.215)
Robust SE (0.226) (0.304) (0.238) (0.276) (0.242)
Observations 128 128 128 128 128
R-squared 0.569 0.357 0.645 0.312 0.410

NOTE. — Standard errors in parentheses. We report clustered and robust standard errors, and we report the most conservative
significance level. Clustered standard errors are clustered at the recruiment group level. We also report the randomization inference
p-value after 2,000 stratified clustered resampling. Panel A reports the estimates obtained from ordinary least squares (OLS)
regressions. Panel B reports the estimates obtained from two-stage least squares regression (2SLS) regressions using treatment
assignment indicator as an instrument for attendance. In all the regressions, we add the following variables to control for differences
in survey responses. We add the lagged dependent variable. Additionally, we add a dummy for maternal year of birth between 1978
and 1987; dummy for Hispanic ethnicity; dummy for non-Hispanic black; dummy for single mother; dummy for cohabiting mother;
dummy for income below two times the federal poverty line; dummy for mothers with some postsecondary education; dummy for
male child; age of the child at the date of the recording; and dummy for randomization group (block).

* Significant at the 10% level.

** Significant at the 5% level.
*** Significant at the 1% level.
 TABLE 5
Mediation Analysis Results
Dependent Variable: Conversational Turn Counts
Mediators
ITT + Parental Beliefs
ITT
Error-Ridden Factor Score
(1) (2) (3)
Random assignment to control or invitation arm 0.314** 0.253* 0.163
Clustered SE (0.113) (0.137) (0.147)
Robust SE (0.152) (0.148) (0.145)
Maternal beliefs (error-riden) 0.098
Clustered SE (0.094)
Robust SE (0.085)
Maternal beliefs (factor score) 0.198**
Clustered SE (0.074)
Robust SE (0.099)
Observations 90 90 90
p-value Joint Significance of mediators (clustered)
p-value Joint Significance of mediators (robust)
Control Group Mean -0.089 -0.089 -0.089
Monte Carlo 95% CIs for Indirect Effects (clustered) [-0.043; 0.168] [0.026; 0.288]
Monte Carlo 95% CIs for Indirect Effects (robust) [-0.035; 0.169] [0.002; 0.321]

NOTE. — Standard errors in parentheses. We report clustered and robust standard errors, and we
report the most conservative significance level. Clustered standard errors are clustered at the
recruiment group level. In all the regressions, we add the following variables to control for differences in
survey responses. We add the lagged dependent variable. Additionally, we add a dummy for maternal
year of birth between 1978 and 1987; dummy for Hispanic ethnicity; dummy for non-Hispanic black;
dummy for single mother; dummy for cohabiting mother; dummy for income below two times the federal
poverty line; dummy for mothers with some postsecondary education; dummy for male child; age of the
child at the date of the recording; and dummy for randomization group (block). The numbers in square
brackets show the upper limits and lower limits of the Monte Carlo 95% confidence intervals for the
indirect effects of the LENA Start program on the conversational turns through the respective
mediators.

* Significant at the 10% level.

** Significant at the 5% level.
*** Significant at the 1% level.